Using Language Models and Topic Models for XML Retrieval

This paper exposes the results of our participation in the INEX 2007 ad hoc track. We implemented two different models: a mixture language model and a topic model. For the language model, we focused on the question of how shallow features of text display information in an XML document can be used to enhance retrieval effectiveness. Our language model combined estimates based on element full-text and the compact representation of the element. We also used non-content priors, including the location the element appears in the original document, and the length of the element path, to boost retrieval effectiveness. For the topic model, we looked at a recent statistical model called Latent Dirichlet Allocation[1], and explored how it could be applied to XML retrieval.